ModSCAN：从视觉和语言模态测量大型视觉-语言模型中的刻板印象偏见
密码学与安全
2024-10-10 v1 计算机与社会
摘要
大型视觉-语言模型(LVLMs)在不断发展和广泛应用，但模型中潜在的刻板印象偏见鲜有探讨。在本研究中，我们提出了 pioneering measurement framework，ModSCAN，to SCAN the stereotypical bias within LVLMs from both vision and language Modality。ModSCAN examines stereotypical biases with respect to two typical stereotypical attributes (gender and race) across three kinds of scenarios: occupations, descriptors, and persona traits.我们的发现表明，1)当前流行的LVLMs显示出显著的刻板印象偏见，CogVLM被 identified as the most biased model; 2)这些刻板印象可能源于训练数据集和预训练模型中固有的偏见; 3)使用特定的提示前缀(来自视觉和语言模态)在减少刻板印象偏见方面效果良好。我们相信我们的工作可以作为理解和解决LVLMs中刻板印象偏见的基础。
引用
@article{arxiv.2410.06967,
title = {$\texttt{ModSCAN}$: Measuring Stereotypical Bias in Large Vision-Language Models from Vision and Language Modalities},
author = {Yukun Jiang and Zheng Li and Xinyue Shen and Yugeng Liu and Michael Backes and Yang Zhang},
journal= {arXiv preprint arXiv:2410.06967},
year = {2024}
}
备注
Accepted in EMNLP 2024. 29 pages, 22 figures